Machine learning guided design of functional materials with targeted properties

Abstract

The impact of machine learning (ML) methods is growing rapidly in materials science. One of the areas where we have shown promise is in the application of ML methods to sequentially guide experiments or computations towards targeted regions in the design space. Here, the need for ML methods is motivated by the fact that the search space is vast and complex. Further, it is intractable for the state-of-the-art high-throughput experimental or computational approaches to navigate the space in an efficient manner. In this review, two examples are discussed that demonstrate the efficacy of classification learning based ML methods in search for new and previously unexplored materials with desired properties. In the first study, ML methods guide density functional theory (DFT) phonon calculations in the design of Ruddlesden-Popper oxides with non-centrosymmetric (NCS) structures. Novel NCS compounds were predicted, whose thermodynamic stabilities were evaluated by DFT convex hull calculations. In the second study, ML methods enabled rapid screening of xBi[Mey′Me(1-y)″]O3–(1–x)PbTiO3 compositions in the desired perovskite crystal structure, where Me′ and Me″ are two octahedral site cations. The predictions were validated by experiments and new {Me′,Me″} cation pairs were identified that also have high ferroelectric Curie temperature.